From de35a8626267e7ec63ce13e8a45c42c1a67b096f Mon Sep 17 00:00:00 2001
From: Arun Sharma <arun@sharma-home.net>
Date: Wed, 30 Sep 2026 18:05:15 -0700
Subject: [PATCH] fix(connection): segregate implicit prepare cache by full
 list element types

_pybind_value_signature keyed list params on the first element only,
while C++ prepares the max logical type over ALL elements. Calls whose
first row was narrow reused a stale narrower plan, silently dropping
rows or truncating values (300 bound as 44, Refs #57). Aggregate
element signatures across the whole list (max int width, merged struct
fields), mirroring C++ tryGetMaxLogicalType; homogeneous int lists
still key as INT64 like pyHomogeneousListType.
---
 src_py/connection.py                       | 56 ++++++++++++++++++++--
 test/test_pybind_implicit_prepare_cache.py | 21 ++++++++
 2 files changed, 72 insertions(+), 5 deletions(-)

diff --git a/src_py/connection.py b/src_py/connection.py
index 8ca77b3..b528195 100644
--- a/src_py/connection.py
+++ b/src_py/connection.py
@@ -247,12 +247,58 @@ def _pybind_value_signature(value: Any) -> tuple:
             tuple((str(k), _pybind_value_signature(v)) for k, v in items),
         )
     if isinstance(value, (list, tuple)):
-        homogeneous = _pybind_homogeneous_list_signature(value)
-        if homogeneous is not None:
-            return homogeneous
-        if not value:
+        int_width_order = {
+            "int8": 0,
+            "uint8": 1,
+            "int16": 2,
+            "uint16": 3,
+            "int32": 4,
+            "uint32": 5,
+            "int64": 6,
+        }
+
+        def _merge(a: tuple, b: tuple) -> tuple:
+            # Any/unknown poisoning: fall back to first-seen (keeps old key).
+            if a == b:
+                return a
+            if a == ("any",) or b == ("any",):
+                return ("any",)
+            if (
+                len(a) == 1
+                and len(b) == 1
+                and a[0] in int_width_order
+                and b[0] in int_width_order
+            ):
+                # Mirror C++ tryGetMaxLogicalType over list children: the
+                # prepared type is the max over ALL elements, so the cache key
+                # must reflect the widest int, not the first element's width.
+                return a if int_width_order[a[0]] >= int_width_order[b[0]] else b
+            if a[0] == "list" and b[0] == "list" and len(a) == 2 and len(b) == 2:
+                return ("list", _merge(a[1], b[1]))
+            if a[0] == "struct" and b[0] == "struct":
+                fa, fb = dict(a[1]), dict(b[1])
+                if set(fa) != set(fb):
+                    return ("struct", tuple(sorted(set(fa) | set(fb))))
+                return ("struct", tuple((k, _merge(fa[k], fb[k])) for k in fa))
+            if a[0] == "map" and b[0] == "map":
+                return ("map", _merge(a[1], b[1]), _merge(a[2], b[2]))
+            # Heterogeneous / unmergeable: keep first-seen (old behaviour).
+            return a
+
+        merged: tuple | None = None
+        for item in value:
+            if item is None:
+                continue
+            sig = _pybind_value_signature(item)
+            merged = sig if merged is None else _merge(merged, sig)
+        if merged is None:
             return ("list", ("any",))
-        return ("list", _pybind_value_signature(value[0]))
+        # Mirror C++ pyHomogeneousListType: a homogeneous int/bool/float list
+        # prepares as LIST(INT64)/LIST(BOOL)/LIST(DOUBLE) regardless of value
+        # widths, so normalise the key the same way to avoid duplicate prepares.
+        if len(merged) == 1 and merged[0] in int_width_order:
+            return ("list", ("int64",))
+        return ("list", merged)
     return ("unknown", type(value).__name__)
 
 
diff --git a/test/test_pybind_implicit_prepare_cache.py b/test/test_pybind_implicit_prepare_cache.py
index e0466ca..1916580 100644
--- a/test/test_pybind_implicit_prepare_cache.py
+++ b/test/test_pybind_implicit_prepare_cache.py
@@ -220,3 +220,24 @@ def test_pybind_implicit_prepare_reuses_same_struct_shape(
         {"st": {"key1": [1, 2, 3], "value": [3, 7, 98]}},
         {"st": {"key1": [4, 5, 6], "value": [7, 8, 9]}},
     ]
+
+
+def test_pybind_implicit_prepare_segregates_struct_list_by_int_width(
+    fake_pybind_connection: _FakePybindConnection,
+) -> None:
+    """
+    Regression test for issue #57.
+
+    A list-of-struct parameter whose later rows need a wider int type than
+    the first call's must not reuse the narrower cached statement (values
+    were silently dropped/truncated through the stale plan).
+    """
+    conn = lb.Connection(_FakeDatabase())
+
+    conn.execute("UNWIND $rows AS r RETURN r.d", {"rows": [{"d": 0}]})
+    conn.execute("UNWIND $rows AS r RETURN r.d", {"rows": [{"d": 1}, {"d": 128}]})
+    conn.execute("UNWIND $rows AS r RETURN r.d", {"rows": [{"d": 2}, {"d": 128}]})
+    conn.execute("UNWIND $rows AS r RETURN r.d", {"rows": [{"d": 4}, {"d": 300}]})
+
+    # int8 -> uint8 -> uint8 (reuse) -> int16 : 3 prepared statements.
+    assert len(fake_pybind_connection.prepare_calls) == 3